When you use Claude Code or Codex, the model itself is only one piece. The rest — the execution loop, context management, error recovery, and session persistence — is the harness. It's the layer of code that sits between you and the LLM, turning a stateless model into something that can work on real tasks across many turns and tool calls. It's what makes Claude Code so incredibly powerful.

Developers use agent harnesses every day, but most treat them as black boxes. You type a prompt, tools run, code appears. What happens between your message and the model's response is more complex and more fragile than people assume.

This talk opens up that systems layer. We'll look at how today's popular harnesses solve the same hard problems: keeping conversations alive when they outgrow the context window, recovering from API failures mid-stream, isolating sub-agent work so the parent doesn't drown in noise, and persisting state across sessions that span hours or days.

The talk includes all the learnings from analyzing the source code of popular harnesses, as well as building OpenHarness, an open-source SDK for building powerful agent harnesses.